Run trtexec on yolo v11 model variations.

trtexec_v11.sh

# sudo apt-get install -y cuda-toolkit-<major>-<minor>
# sudo apt-get install -y tensorrt nvidia-tensorrt-dev
# sudo nvpmodel -m 0
# sudo jetson_clocks

out_file="/home/user/yolo11_trtexec.txt"
rm ${out_file}

variants=("yolo11n" "yolo11s" "yolo11m" "yolo11l" "yolo11x")
image_sizes=("320" "384" "416" "512" "608" "640")

for variant in "${variants[@]}"; do
  for size in "${image_sizes[@]}"; do
    echo -e "------ ${variant} ${size} ------" >> ${out_file} 2>&1
    /usr/src/tensorrt/bin/trtexec \
        --onnx=./models/${variant}.onnx \
        --fp16 --inputIOFormats=fp16:chw \
        --saveEngine=./models/${variant}_${size}_fp_16.engine \
        --shapes=\'images\':1x3x${size}x${size} \
        --noDataTransfers --useSpinWait \
        --timingCacheFile=./timing_cache \
        --iterations=100 --duration=5 2>&1 | tee -a ${out_file}
  done
done


# With dump profile (per layer info), add --dumpProfile, e.g.
# /usr/src/tensorrt/bin/trtexec \
#     --onnx=./models/${variant}.onnx \
#     --fp16 --inputIOFormats=fp16:chw \
#     --shapes=\'images\':1x3x${size}x${size} \
#     --noDataTransfers --useSpinWait \
#     --timingCacheFile=./timing_cache \
#     --iterations=100 --duration=5 \
#     --dumpProfile 2>&1 | tee -a ${out_file}